We evaluate multimodal large language models (MLLMs) for topic-aligned captioning in financial short-form videos (SVs) by testing joint reasoning over transcripts (T), audio (A), and video (V). Using 624 annotated YouTube SVs, we assess all seven modality combinations (T, A, V, TA, TV, AV, TAV) across five topics: main recommendation, sentiment analysis, video purpose, visual analysis, and financial entity recognition. Video alone performs strongly on four of five topics, underscoring its value for capturing visual context and effective cues such as emotions, gestures, and body language. Selective pairs such as TV or AV often surpass TAV, implying that too many modalities may introduce noise. These results establish the first baselines for financial short-form video captioning and illustrate the potential and challenges of grounding complex visual cues in this domain. All code and data can be found on our Github under the CC-BY-NC-SA 4.0 license.

Recently, advancements in AI counseling based on large language models have shown significant progress. However, existing studies employ a one-time generation approach to synthesize multi-turn dialogue samples, resulting in low therapy fidelity and failing to capture the decision-making rationale behind each response. In this work, we propose CATCH, a novel data synthesis framework designed to address these challenges. Specifically, to improve therapy fidelity, we introduce the Progressive Dialogue Synthesis strategy, which extracts goals, resources, and solutions from a client's self-report, organizes them into structured outlines, and then incrementally generates stage-aligned counseling dialogues. To capture decision-making rationale behind each response, we propose the Memory-Driven Dynamic Planning thinking pattern that integrates memory enhancement, global planning, and strategy reasoning; a collaborative multi-agent optimizer then leverages MDP to attach explicit chain-of-thought to each dialogue turn. Extensive experiments and human evaluations demonstrate that CATCH significantly enhances fidelity and logical coherence in AI counseling.

Text anonymization is essential for responsibly developing and deploying AI in high-stakes domains such as healthcare, social services, and law. In this work, we propose a novel methodology for privacy-preserving synthetic text generation that leverages the principles of de-identification and the Hiding In Plain Sight (HIPS) theory. Our approach introduces entity-aware control codes to guide controllable generation using either in-context learning (ICL) or prefix tuning. The ICL variant ensures privacy levels consistent with the underlying de-identification system, while the prefix tuning variant incorporates a custom masking strategy and loss function to support scalable, high-quality generation. Experiments on legal and clinical datasets demonstrate that our method achieves a strong balance between privacy protection and utility, offering a practical and effective solution for synthetic text generation in sensitive domains.

Current search techniques are limited to standard RAG query-document applications. In this paper, we propose a novel technique to expand the code and index for predicting the required APIs, directly enabling high-quality, end-to-end code generation for auto-completion and agentic AI applications. We address the problem of API leaks in current code-to-code benchmark datasets by introducing a new dataset built from real-world ServiceNow Script Includes that capture the challenge of unclear API usage intent in the code. Our evaluation metrics show that this method achieves 87.86% top-40 retrieval accuracy, allowing the critical context with APIs needed for successful downstream code generation. To enable real-time predictions, we develop a comprehensive post-training pipeline that optimizes a compact 0.6B reranker through synthetic dataset generation, supervised fine-tuning, and reinforcement learning. This approach enables our compact reranker to outperform a much larger 8B model while maintaining 2.5x reduced latency, effectively addressing the nuances of enterprise-specific code without the computational overhead of larger models.

The Mixture of Experts (MoE) architecture is an important method for scaling Large Language Models (LLMs). It increases model capacity while keeping computation cost low. However, the ultra-large MoE models still have hundreds of billions of parameters, requiring massive memory/storage and leading to difficulties for deployment on resource-constrained edge platforms. Pruning or quantization alone can hardly address the issue, because of the super-aggressive compression ratio with significantly degraded accuracy and output quality. To facilitate the deployment of ultra-large MoEs on edge platforms, we propose a collaborative compression framework by combining expert pruning, mixed-precision quantization, and activation optimization. It can effectively reduce the storage footprint of the ultra-large MoE DeepSeek-V3 from 1.3TB to 103GB, while preserving high output quality with better accuracy than traditional uniform low-bit quantization methods. To the best of our knowledge, we are the first to deploy a compressed model from the ultra-large DeepSeek-V3 on the platform with a strict 128GB total memory limit. Our comprehensive experiments on multiple benchmarks under various memory constraints demonstrate the effectiveness of our method with smaller model sizes and higher accuracy than uniform low-bit quantization methods.

Benchmarks shape scientific conclusions about model capabilities and steer model development. This creates a feedback loop: stronger benchmarks drive better models, and better models demand more discriminative benchmarks. Ensuring benchmark reliability is therefore essential for trustworthy evaluation and meaningful progress. In this work, we study benchmark reliability from a distributional perspective and introduce benchmark harmony, which measures how uniformly a model's performance is distributed across the subdomains of a benchmark. We posit that high harmony is a desirable benchmark property, indicating that the aggregate metric reflects uniform competence across subdomains. Across 19 multiple-choice benchmarks and five model families, we map each benchmark onto a mean-variance plane of harmony computed across models, where high mean and low variance signal more reliable evaluation. Our analysis shows that less harmonious benchmarks can give misleading results, since overall accuracy may be disproportionately influenced by specific subdomains. For instance, ARC-Easy is overwhelmed by questions on Biological Concepts, overshadowing other critical subdomains such as Geography, Physics, Chemistry, and Environmental Science. By recommending that harmony should be reported alongside accuracy, we reframe evaluation from simple performance averages to a more robust, distributionally reliable measurement of performance.

We propose a method to improve Visual Question Answering (VQA) with Retrieval-Augmented Generation (RAG) by introducing text-grounded object localization. Rather than retrieving information based on the entire image, our approach enables the model to generate a bounding box around the object most relevant to the question, allowing for targeted image cropping and focused retrieval. This reduces background noise, improves alignment between visual and textual cues, and helps mitigate hallucinations. Our RAG method enhances context-aware VQA responses increased the accuracy from 22.19% to 25.64%, with an absolute increase of 3.45 percentage points, compared to the baseline Llama-3.2-Vision-11B agent. We also proposed a de-hallucination method based on question type which can effectively reduce the hallucination rate from 65.79% to 13.88% and improves the truthfulness score.

Current online reinforcement learning (RL) algorithms like GRPO share a key limitation in LLM reasoning: they cannot learn from problems that are "unsolvable" to the model. In other words, they can only improve performance on problems where the model is capable of exploring the correct answer. Consequently, the model's "upper limit" remains unchanged after RL training, even though the likelihood of solving easier, solvable problems may increase. These hard samples cannot contribute to training, as no rollouts yield rewards and thus no gradients are produced. To unlock learning from these hard samples, we propose NuRL, a "nudging" method that aims to push the upper bound of LLM reasoning using self-generated hints, i.e., abstract cues that help reduce the problem difficulty for the model. Given a question and its gold answer, the model generates a CoT and then produces a hint containing the core knowledge needed to solve the problem. During training, we generate G rollouts from the base policy and use the pass rate to decide whether the hint should be injected. For hard samples with a 0% pass rate, we inject the hint and regenerate a new batch of trajectories. This yields two benefits: (1) the hint boosts pass rates (from 0% to non-zero), thereby introducing training signals for previously unsolvable samples, and (2) the hints are self-generated, avoiding distributional shift and do not rely on external models. NuRL achieves consistent improvements across 6 benchmarks and 3 models, while remaining complementary to test-time scaling. Notably, NuRL can raise the model's upper limit, whereas GRPO leaves pass@1024 unchanged from the base model. Furthermore, we present a systematic study of what makes an effective hint and when hints are most useful. Interestingly, the best hints are abstract and high-level, and are most beneficial when applied necessarily and after GRPO has converged.

Mainstream news organizations shape public perception not only directly through the articles they publish but also through the choices they make about which topics to cover (or ignore) and how to frame the issues they do decide to cover. However, measuring these subtle forms of media bias at scale remains a challenge. Here, we introduce a large, ongoing (from January 1, 2024 to present), near real-time dataset and computational framework developed to enable systematic study of selection and framing bias in news coverage. Our pipeline integrates large language models (LLMs) with scalable, near-real-time news scraping to extract structured annotations -- including political lean, tone, topics, article type, and major events -- across hundreds of articles per day. We quantify these dimensions of coverage at multiple levels -- the sentence level, the article level, and the publisher level -- expanding the ways in which researchers can analyze media bias in the modern news landscape. In addition to a curated dataset, we also release an interactive web platform for convenient exploration of these data. Together, these contributions establish a reusable methodology for studying media bias at scale, providing empirical resources for future research. Leveraging the breadth of the corpus over time and across publishers, we also present some examples (focused on the 150,000+ articles examined in 2024) that illustrate how this novel data set can reveal insightful patterns in news coverage and bias, supporting academic research and real-world efforts to improve media accountability.

Diffusion large language models (dLLMs) have shown great potential in large-scale language modeling, and there is an increasing interest in further improving the capacity to solve complex problems by guiding the reasoning process step by step. Common practice for autoregressive language models typically learns a process reward model with dense annotation for each intermediate step. However, this is challenging for dLLMs where the generation is in an any-order fashion and intermediate states are partially masked sentences. To this end, in this paper, we propose reward-free guidance (RFG), a principled method for guiding the reasoning trajectory of dLLMs without explicit process reward. The key idea of RFG is to parameterize the process reward by log-likelihood ratios of the enhanced and reference dLLMs, where the enhanced model can be easily obtained by any off-the-shelf dLLM that has been post-trained with reinforcement learning (RL) or supervised fine-tuning (SFT). We provide theoretical justification that RFG induces the reward-guided sampling distribution with no additional reward. We conduct comprehensive experiments on four challenging mathematical reasoning and code generation benchmarks using a diverse suite of dLLMs enhanced with various post-training methods. RFG consistently yields significant improvements across all tasks and model types, achieving accuracy gains of up to 9.2%. These findings establish RFG as a general training-free framework that scales test-time reasoning without reliance on external reward models. By scaling up mask-predict pretraining on large-scale corpora through bidirectional computation, dLLMs have shown surprisingly competitive or even superior performance over autoregressive (AR) model baselines (Prabhudesai et al., 2025). Despite the impressive advancements, the current success of dLLMs is primarily limited to pre-training or continue-training on a specific domain, with limited exploration in test-time computation and alignment.